LLM Rubrics Boost Clinical Supervised Learning

A recent preprint on arXiv (2603.11679) introduces an agentic pipeline that enables large language models (LLMs) to create both global and local rubrics, enhancing supervised learning for intricate multimodal datasets. The global rubric is a programmatic specification derived from a varied collection of text-serialized examples, aimed at standardizing input formats. In contrast, local rubrics serve as interpretive summaries tailored to specific tasks. When evaluated on 15 clinical tasks from the EHRSHOT benchmark, the rubric methods surpassed count-feature models, basic LLM baselines, and a clinical foundation model that had been pretrained on similar datasets. This approach minimizes the necessity for domain expertise in designing input representations.